Torsion elastic constants obtained from fluorescence polarization anisotropy (FPA) measurements on fifty-three 24 and 27 base-pair (bp) DNAs were recently reported [F. Pedone, F. Mazzei, D. Santoni, Sequence-dependent DNA torsional rigidity: a tetranucleotide code, Biophys. Chem. 112 (2004) 77-88; F. Pedone, F. Mazzei, M. Matzeu, F. Barone, Torsional constant of 27-mer DNA oligomers of different sequences, Biophys. Chem. 94 (2001) 175-184]. The problem of extracting reliable torsion elastic constants (alpha) from FPA measurements on such short DNAs is examined in detail. The difficulty is illustrated by two (fictitious) 24 bp DNAs with approximately 5-fold different torsion elastic constants and 10% different initial anisotropies (r(0)), which exhibit practically indistinguishable anisotropy decays for all t>1 ns. FPA data were simulated for 24 bp DNAs with different input values of alpha and r(0) in the presence and absence of Poisson noise, and were fitted using different choices of the adjustable and fixed parameters. Experimental data for a 24 bp DNA were fitted in a similar manner. For either the simulated or experimental FPA data, it was not possible to determine both the initial anisotropy, r(0), and the torsion elastic constant, alpha, in a reliable (i.e. statistically significant) manner in the presence of Poisson noise. When r(0) is assumed to be fixed at any particular value in the fitting protocol, a unique best-fit value of alpha is obtained, but that best-fit alpha is extremely sensitive to small deviations of the assumed fixed value of r(0) away from the input r(0)-value of the simulated data. Pedone et al. fitted their FPA data by assuming that r(0)=0.360, and adjusting alpha, the hydrodynamic radius (R(H)), and effective length (L). In fact, the reported best-fit values of R(H) and L lay significantly outside their expected ranges. When this same fitting protocol is applied to simulated data for 27 bp DNAs, better overall agreement with the reported experimental values (alpha, R(H), and L) is obtained for a model, wherein all DNAs have the same typical input alpha=5.9 x 10(-12) dyn cm, R(H)=10.0 A, and L=27 (3.4)+2.7=94.5 A, but a 1.00- to 1.13-fold range of r(0)-values, than for the model of Pedone et al., wherein all DNAs have the same input r(0)=0.360, R(H)=10.0 A, and L=94.5 A, but a approximately 3-fold range of alpha-values. It is concluded that, in the absence of reliable independent estimates of r(0) for every DNA, the alpha-values reported for 24 and 27 bp DNAs cannot be regarded as experimentally justified. The reliability of the torsion elastic constants reported for the 136 distinct tetranucleotide steps, which are inferred from the values reported for the fifty-three 24 and 27 bp DNAs, is also briefly discussed.